LLaMA.cpp releases version b9729 with binaries for macOS, Linux, Android, Windows, and openEuler across multiple architectures. The release includes CPU, Vulkan, OpenVINO, SYCL, and ROCm support, along with a new UI package. Internal references to 'webui' have been removed.
Users seeking optimal llama.cpp settings for gemma 4 models on an AMD GPU with 16GB VRAM ask whether trial and error is necessary. They reference Google's default settings for temperature, top-p, and top-k but note inconsistent results, indicating a need for more targeted guidance beyond official documentation.
A long-context decode performance cliff on AMD Radeon AI PRO R9700 (RDNA4) was resolved by enabling AITER Unified Attention in vLLM 0.22.1. The fix involves relaxing a CDNA gate to include RDNA4, disabling other attention backends, and using bf16 KV cache, resulting in significant speedups across all context lengths. FP8 KV is ineffective on this hardware, and the model's native 262K context is fully achievable with bf16, offering ~2.9× concurrency without needing FP8.
LLaMA.cpp version b9728 introduces support for comment lines in --api-key-file configuration. The release includes pre-built binaries for macOS, Linux, Android, Windows, and openEuler across multiple architectures and hardware acceleration options, including Vulkan, CUDA, OpenVINO, and SYCL.
Users report that tools like LM Studio and Odysseus are limited by search engine request caps, often at 10 per day or hour, without API access. They suggest creating DuckDuckGo API accounts for better search access, but note that frontends rarely prompt for this. The post asks whether Hermes or Pi offer improved solutions.
A user reports their CPU and RAM are reaching 100% during simple test prompts, while the GPU is underutilized. They question whether their RTX 3050 8GB GPU can run Quen3.5:9b locally, noting that in theory it should be feasible.
llama.cpp version b9718 consolidates slot selection into a single function, get_available_slot, while maintaining LCP similarity checks for prompt cache updates. The release includes binary builds for macOS, Linux, Android, Windows, and openEuler across multiple architectures and hardware acceleration options.
llama.cpp has released version b9721, offering binaries for macOS, Linux, Android, Windows, and openEuler across various architectures. The release includes CPU, Vulkan, ROCm, OpenVINO, SYCL, and HIP support, with a dedicated UI package. A feature for Apple Silicon with KleidiAI is currently disabled.
GLM-5.2, the strongest open model to date, can now run locally using llama.cpp and Unsloth Studio. The 2-bit quantized model retains ~82% accuracy after reducing size from 1.51TB to 238GB, a 84% reduction, and is compatible with 256GB RAM or VRAM setups.
LLaMA.cpp version b9715 introduces CUDA support for GGML_OP_COL2IM_1D, building on a CPU implementation. The release includes binaries for macOS, Linux, Android, Windows, and openEuler across multiple architectures and acceleration frameworks, including Vulkan, ROCm, OpenVINO, and SYCL.
UltraQuant enables 4-bit KV caching for context-heavy agents, reducing P50 time-to-first-token by 3.47x in late rounds and boosting output throughput by 1.63x over FP8 KV baseline. It achieves this using FP8 queries, FP4 KV tensors, UE8M0 group scales, and native scaled-MFMA on AMD CDNA4 GPUs, with optimizations for decode-attention kernels and robust design choices like asymmetric K/V treatment and Walsh-Hadamard rotation.
HEPTv2 achieves 98.6% tracking efficiency with 0.8% fake rate on TrackML, using only 15 ms inference time and 0.4 GB memory per event. It outperforms prior transformer and graph-based methods in efficiency and reduces latency by factors of 7 and 38–52, respectively, enabling real-time particle reconstruction at the HL-LHC.
UltraQuant introduces a 4-bit KV caching method tailored for context-heavy agent workloads. It achieves 3.47x reduction in P50 time-to-first-token in late rounds and 1.63x higher output throughput compared to FP8 KV caching, using FP8 queries, FP4 KV tensors, and native AMD CDNA4 scaled-MFMA support.
Execution-state capsules enable graph-bound checkpointing and restoration of complete execution state, including KV, recurrent, and convolution states, for low-latency, small-batch on-device AI serving. On RTX 5090 and Jetson AGX Thor, capsule restore achieves byte-exact and token-identical correctness, with sub-millisecond GPU operations and TTFT speedups up to 27x at 16k tokens, demonstrating significant latency reduction in interactive AI workflows.
StreamKL introduces a fused GPU primitive that eliminates quadratic memory usage in attention distillation by streaming query-key tiles through on-chip SRAM. It achieves up to 43x speedup in forward and 14x in backward passes, reducing extra HBM footprint from O(N_QN_K) to O(1), enabling long-context distillation on a single GPU.
An adaptive-trunk DeepONet framework predicts localized structural responses in long-span bridges with high accuracy. By using distance-aware features and a stiffness-informed Schur complement, it achieves FEM-level accuracy with less than 5% error, reducing total response evaluation time by 60x and inference speed by up to four orders of magnitude compared to finite element methods.
Sevra, a serving-layer controller, selectively verifies answers to improve accuracy and reduce token usage. On \mathfive, it achieves 76.3% accuracy with 26.8% fewer post-generation tokens and halved harmful flips, while on \gsm it verifies only 3.0% of examples, boosting accuracy to 94.5% and cutting verification tokens by 91.2%. The study shows that initial solve length and explicit control needs determine optimal verification strategy.
This paper introduces a four-layer technical architecture for token-oriented inference optimization, including Multi-model Fusion, Model Optimization, Compute-Model Fusion, and Compute-Network-Model Fusion. It reviews key technologies and industry status, analyzing their real-world application value in reducing token costs, enhancing service efficiency, and ensuring stable token supply.
GLM-5.2 UD-IQ2_M runs at ~7.3 tokens per second on 4×RTX 3090s with 192GB DDR5 RAM using llama.cpp expert offload. Reducing quantization from IQ2 to IQ1 provided no speed gain, while increasing CPU threads from 6 to 12 improved performance by 22%. Decode is limited by CPU compute, not memory bandwidth, and the offloaded experts must be explicitly distributed across GPUs to avoid out-of-memory errors.
DiffusionGemma 26B runs at up to 475t/s on a 4090 via vLLM with INT4 AWQ quantization, achieving speeds between 290t/s and 700t/s based on output length. However, it suffers from single-user operation, lower response accuracy, rapid context loss, and slower time-to-first-token compared to standard 26B models.